1 //===- bolt/Profile/BoltAddressTranslation.cpp ----------------------------===// 2 // 3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. 4 // See https://llvm.org/LICENSE.txt for license information. 5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception 6 // 7 //===----------------------------------------------------------------------===// 8 9 #include "bolt/Profile/BoltAddressTranslation.h" 10 #include "bolt/Core/BinaryFunction.h" 11 #include "llvm/Support/DataExtractor.h" 12 #include "llvm/Support/Errc.h" 13 14 #define DEBUG_TYPE "bolt-bat" 15 16 namespace llvm { 17 namespace bolt { 18 19 const char *BoltAddressTranslation::SECTION_NAME = ".note.bolt_bat"; 20 21 void BoltAddressTranslation::writeEntriesForBB(MapTy &Map, 22 const BinaryBasicBlock &BB, 23 uint64_t FuncAddress) { 24 const uint64_t BBOutputOffset = 25 BB.getOutputAddressRange().first - FuncAddress; 26 const uint32_t BBInputOffset = BB.getInputOffset(); 27 28 // Every output BB must track back to an input BB for profile collection 29 // in bolted binaries. If we are missing an offset, it means this block was 30 // created by a pass. We will skip writing any entries for it, and this means 31 // any traffic happening in this block will map to the previous block in the 32 // layout. This covers the case where an input basic block is split into two, 33 // and the second one lacks any offset. 34 if (BBInputOffset == BinaryBasicBlock::INVALID_OFFSET) 35 return; 36 37 LLVM_DEBUG(dbgs() << "BB " << BB.getName() << "\n"); 38 LLVM_DEBUG(dbgs() << " Key: " << Twine::utohexstr(BBOutputOffset) 39 << " Val: " << Twine::utohexstr(BBInputOffset) << "\n"); 40 // In case of conflicts (same Key mapping to different Vals), the last 41 // update takes precedence. Of course it is not ideal to have conflicts and 42 // those happen when we have an empty BB that either contained only 43 // NOPs or a jump to the next block (successor). Either way, the successor 44 // and this deleted block will both share the same output address (the same 45 // key), and we need to map back. We choose here to privilege the successor by 46 // allowing it to overwrite the previously inserted key in the map. 47 Map[BBOutputOffset] = BBInputOffset; 48 49 for (const auto &IOPair : BB.getOffsetTranslationTable()) { 50 const uint64_t OutputOffset = IOPair.first + BBOutputOffset; 51 const uint32_t InputOffset = IOPair.second; 52 53 // Is this the first instruction in the BB? No need to duplicate the entry. 54 if (OutputOffset == BBOutputOffset) 55 continue; 56 57 LLVM_DEBUG(dbgs() << " Key: " << Twine::utohexstr(OutputOffset) << " Val: " 58 << Twine::utohexstr(InputOffset) << " (branch)\n"); 59 Map.insert( 60 std::pair<uint32_t, uint32_t>(OutputOffset, InputOffset | BRANCHENTRY)); 61 } 62 } 63 64 void BoltAddressTranslation::write(const BinaryContext &BC, raw_ostream &OS) { 65 LLVM_DEBUG(dbgs() << "BOLT-DEBUG: Writing BOLT Address Translation Tables\n"); 66 for (auto &BFI : BC.getBinaryFunctions()) { 67 const BinaryFunction &Function = BFI.second; 68 // We don't need a translation table if the body of the function hasn't 69 // changed 70 if (Function.isIgnored() || (!BC.HasRelocations && !Function.isSimple())) 71 continue; 72 73 LLVM_DEBUG(dbgs() << "Function name: " << Function.getPrintName() << "\n"); 74 LLVM_DEBUG(dbgs() << " Address reference: 0x" 75 << Twine::utohexstr(Function.getOutputAddress()) << "\n"); 76 MapTy Map; 77 const bool IsSplit = Function.isSplit(); 78 for (const BinaryBasicBlock *const BB : Function.getLayout().blocks()) { 79 if (IsSplit && BB->isCold()) 80 break; 81 writeEntriesForBB(Map, *BB, Function.getOutputAddress()); 82 } 83 Maps.insert(std::pair<uint64_t, MapTy>(Function.getOutputAddress(), Map)); 84 85 if (!IsSplit) 86 continue; 87 88 // Cold map 89 Map.clear(); 90 LLVM_DEBUG(dbgs() << " Cold part\n"); 91 for (const BinaryBasicBlock *const BB : Function.getLayout().blocks()) { 92 if (!BB->isCold()) 93 continue; 94 writeEntriesForBB(Map, *BB, Function.cold().getAddress()); 95 } 96 Maps.insert(std::pair<uint64_t, MapTy>(Function.cold().getAddress(), Map)); 97 ColdPartSource.insert(std::pair<uint64_t, uint64_t>( 98 Function.cold().getAddress(), Function.getOutputAddress())); 99 } 100 101 const uint32_t NumFuncs = Maps.size(); 102 OS.write(reinterpret_cast<const char *>(&NumFuncs), 4); 103 LLVM_DEBUG(dbgs() << "Writing " << NumFuncs << " functions for BAT.\n"); 104 for (auto &MapEntry : Maps) { 105 const uint64_t Address = MapEntry.first; 106 MapTy &Map = MapEntry.second; 107 const uint32_t NumEntries = Map.size(); 108 LLVM_DEBUG(dbgs() << "Writing " << NumEntries << " entries for 0x" 109 << Twine::utohexstr(Address) << ".\n"); 110 OS.write(reinterpret_cast<const char *>(&Address), 8); 111 OS.write(reinterpret_cast<const char *>(&NumEntries), 4); 112 for (std::pair<const uint32_t, uint32_t> &KeyVal : Map) { 113 OS.write(reinterpret_cast<const char *>(&KeyVal.first), 4); 114 OS.write(reinterpret_cast<const char *>(&KeyVal.second), 4); 115 } 116 } 117 const uint32_t NumColdEntries = ColdPartSource.size(); 118 LLVM_DEBUG(dbgs() << "Writing " << NumColdEntries 119 << " cold part mappings.\n"); 120 OS.write(reinterpret_cast<const char *>(&NumColdEntries), 4); 121 for (std::pair<const uint64_t, uint64_t> &ColdEntry : ColdPartSource) { 122 OS.write(reinterpret_cast<const char *>(&ColdEntry.first), 8); 123 OS.write(reinterpret_cast<const char *>(&ColdEntry.second), 8); 124 LLVM_DEBUG(dbgs() << " " << Twine::utohexstr(ColdEntry.first) << " -> " 125 << Twine::utohexstr(ColdEntry.second) << "\n"); 126 } 127 128 outs() << "BOLT-INFO: Wrote " << Maps.size() << " BAT maps\n"; 129 outs() << "BOLT-INFO: Wrote " << NumColdEntries 130 << " BAT cold-to-hot entries\n"; 131 } 132 133 std::error_code BoltAddressTranslation::parse(StringRef Buf) { 134 DataExtractor DE = DataExtractor(Buf, true, 8); 135 uint64_t Offset = 0; 136 if (Buf.size() < 12) 137 return make_error_code(llvm::errc::io_error); 138 139 const uint32_t NameSz = DE.getU32(&Offset); 140 const uint32_t DescSz = DE.getU32(&Offset); 141 const uint32_t Type = DE.getU32(&Offset); 142 143 if (Type != BinarySection::NT_BOLT_BAT || 144 Buf.size() + Offset < alignTo(NameSz, 4) + DescSz) 145 return make_error_code(llvm::errc::io_error); 146 147 StringRef Name = Buf.slice(Offset, Offset + NameSz); 148 Offset = alignTo(Offset + NameSz, 4); 149 if (Name.substr(0, 4) != "BOLT") 150 return make_error_code(llvm::errc::io_error); 151 152 if (Buf.size() - Offset < 4) 153 return make_error_code(llvm::errc::io_error); 154 155 const uint32_t NumFunctions = DE.getU32(&Offset); 156 LLVM_DEBUG(dbgs() << "Parsing " << NumFunctions << " functions\n"); 157 for (uint32_t I = 0; I < NumFunctions; ++I) { 158 if (Buf.size() - Offset < 12) 159 return make_error_code(llvm::errc::io_error); 160 161 const uint64_t Address = DE.getU64(&Offset); 162 const uint32_t NumEntries = DE.getU32(&Offset); 163 MapTy Map; 164 165 LLVM_DEBUG(dbgs() << "Parsing " << NumEntries << " entries for 0x" 166 << Twine::utohexstr(Address) << "\n"); 167 if (Buf.size() - Offset < 8 * NumEntries) 168 return make_error_code(llvm::errc::io_error); 169 for (uint32_t J = 0; J < NumEntries; ++J) { 170 const uint32_t OutputAddr = DE.getU32(&Offset); 171 const uint32_t InputAddr = DE.getU32(&Offset); 172 Map.insert(std::pair<uint32_t, uint32_t>(OutputAddr, InputAddr)); 173 LLVM_DEBUG(dbgs() << Twine::utohexstr(OutputAddr) << " -> " 174 << Twine::utohexstr(InputAddr) << "\n"); 175 } 176 Maps.insert(std::pair<uint64_t, MapTy>(Address, Map)); 177 } 178 179 if (Buf.size() - Offset < 4) 180 return make_error_code(llvm::errc::io_error); 181 182 const uint32_t NumColdEntries = DE.getU32(&Offset); 183 LLVM_DEBUG(dbgs() << "Parsing " << NumColdEntries << " cold part mappings\n"); 184 for (uint32_t I = 0; I < NumColdEntries; ++I) { 185 if (Buf.size() - Offset < 16) 186 return make_error_code(llvm::errc::io_error); 187 const uint32_t ColdAddress = DE.getU64(&Offset); 188 const uint32_t HotAddress = DE.getU64(&Offset); 189 ColdPartSource.insert( 190 std::pair<uint64_t, uint64_t>(ColdAddress, HotAddress)); 191 LLVM_DEBUG(dbgs() << Twine::utohexstr(ColdAddress) << " -> " 192 << Twine::utohexstr(HotAddress) << "\n"); 193 } 194 outs() << "BOLT-INFO: Parsed " << Maps.size() << " BAT entries\n"; 195 outs() << "BOLT-INFO: Parsed " << NumColdEntries 196 << " BAT cold-to-hot entries\n"; 197 198 return std::error_code(); 199 } 200 201 void BoltAddressTranslation::dump(raw_ostream &OS) { 202 const size_t NumTables = Maps.size(); 203 OS << "BAT tables for " << NumTables << " functions:\n"; 204 for (const auto &MapEntry : Maps) { 205 OS << "Function Address: 0x" << Twine::utohexstr(MapEntry.first) << "\n"; 206 OS << "BB mappings:\n"; 207 for (const auto &Entry : MapEntry.second) { 208 const bool IsBranch = Entry.second & BRANCHENTRY; 209 const uint32_t Val = Entry.second & ~BRANCHENTRY; 210 OS << "0x" << Twine::utohexstr(Entry.first) << " -> " 211 << "0x" << Twine::utohexstr(Val); 212 if (IsBranch) 213 OS << " (branch)"; 214 OS << "\n"; 215 } 216 OS << "\n"; 217 } 218 const size_t NumColdParts = ColdPartSource.size(); 219 if (!NumColdParts) 220 return; 221 222 OS << NumColdParts << " cold mappings:\n"; 223 for (const auto &Entry : ColdPartSource) { 224 OS << "0x" << Twine::utohexstr(Entry.first) << " -> " 225 << Twine::utohexstr(Entry.second) << "\n"; 226 } 227 OS << "\n"; 228 } 229 230 uint64_t BoltAddressTranslation::translate(uint64_t FuncAddress, 231 uint64_t Offset, 232 bool IsBranchSrc) const { 233 auto Iter = Maps.find(FuncAddress); 234 if (Iter == Maps.end()) 235 return Offset; 236 237 const MapTy &Map = Iter->second; 238 auto KeyVal = Map.upper_bound(Offset); 239 if (KeyVal == Map.begin()) 240 return Offset; 241 242 --KeyVal; 243 244 const uint32_t Val = KeyVal->second & ~BRANCHENTRY; 245 // Branch source addresses are translated to the first instruction of the 246 // source BB to avoid accounting for modifications BOLT may have made in the 247 // BB regarding deletion/addition of instructions. 248 if (IsBranchSrc) 249 return Val; 250 return Offset - KeyVal->first + Val; 251 } 252 253 Optional<BoltAddressTranslation::FallthroughListTy> 254 BoltAddressTranslation::getFallthroughsInTrace(uint64_t FuncAddress, 255 uint64_t From, 256 uint64_t To) const { 257 SmallVector<std::pair<uint64_t, uint64_t>, 16> Res; 258 259 // Filter out trivial case 260 if (From >= To) 261 return Res; 262 263 From -= FuncAddress; 264 To -= FuncAddress; 265 266 auto Iter = Maps.find(FuncAddress); 267 if (Iter == Maps.end()) 268 return NoneType(); 269 270 const MapTy &Map = Iter->second; 271 auto FromIter = Map.upper_bound(From); 272 if (FromIter == Map.begin()) 273 return Res; 274 // Skip instruction entries, to create fallthroughs we are only interested in 275 // BB boundaries 276 do { 277 if (FromIter == Map.begin()) 278 return Res; 279 --FromIter; 280 } while (FromIter->second & BRANCHENTRY); 281 282 auto ToIter = Map.upper_bound(To); 283 if (ToIter == Map.begin()) 284 return Res; 285 --ToIter; 286 if (FromIter->first >= ToIter->first) 287 return Res; 288 289 for (auto Iter = FromIter; Iter != ToIter;) { 290 const uint32_t Src = Iter->first; 291 if (Iter->second & BRANCHENTRY) { 292 ++Iter; 293 continue; 294 } 295 296 ++Iter; 297 while (Iter->second & BRANCHENTRY && Iter != ToIter) 298 ++Iter; 299 if (Iter->second & BRANCHENTRY) 300 break; 301 Res.emplace_back(Src, Iter->first); 302 } 303 304 return Res; 305 } 306 307 uint64_t BoltAddressTranslation::fetchParentAddress(uint64_t Address) const { 308 auto Iter = ColdPartSource.find(Address); 309 if (Iter == ColdPartSource.end()) 310 return 0; 311 return Iter->second; 312 } 313 314 bool BoltAddressTranslation::enabledFor( 315 llvm::object::ELFObjectFileBase *InputFile) const { 316 for (const SectionRef &Section : InputFile->sections()) { 317 Expected<StringRef> SectionNameOrErr = Section.getName(); 318 if (Error E = SectionNameOrErr.takeError()) 319 continue; 320 321 if (SectionNameOrErr.get() == SECTION_NAME) 322 return true; 323 } 324 return false; 325 } 326 } // namespace bolt 327 } // namespace llvm 328